Deep Learning-Based FDD Massive MIMO-RSMA against Time-Varying Channels

In the massive multiple-input multiple-output (MaMIMO) system with rate-splitting multiple access (RSMA) under the frequency division duplex (FDD) mode, the accurate feedback of channel state information (CSI) is extremely important to conduct the hybrid precoding and improve the spectral efficiency. Deep learning (DL)-based joint optimization of CSI feedback and hybrid precoding can better exploit the potentials of RSMA, while little attention has been devoted to it. To solve this problem, we are the first to jointly design the CSI feedback and hybrid precoding in the MaMIMO-RSMA system under time-varying channels in this paper. Specifically, we propose the CSI rate-splitting network (CsiRSNet), regarding max-min fairness (MMF) as the objective criterion. To address the challenges brought by practical time-varying channels, the long short-term memory (LSTM) modules are introduced to our network for the extraction of temporal correlation among channels. Simulation results prove that the joint optimization can achieve a better MMF rate than existing DL-based methods.